90 likes | 114 Views
E81 CSE 532S: Advanced Multi-Paradigm Software Development. Half-Sync/Half-Async (HSHA) and Leader/Followers (LF) Patterns. Venkita Subramonian, Chris Gill, Nick Haddad, and Steve Donahue Department of Computer Science and Engineering Washington University, St. Louis cdgill@cse.wustl.edu.
E N D
E81 CSE 532S: Advanced Multi-Paradigm Software Development Half-Sync/Half-Async (HSHA) and Leader/Followers (LF) Patterns Venkita Subramonian, Chris Gill, Nick Haddad, and Steve Donahue Department of Computer Science and Engineering Washington University, St. Louis cdgill@cse.wustl.edu
HSHA and LF Patterns • Both are (architectural) concurrency patterns • Both decouple asynchronous, synchronous processing • Synchronous layer may reduce programming complexity and overhead (e.g., fewer context switches, cache misses) • Asynchronous layer improves responsiveness to events • Key differences • HSHA dedicates a thread to handle input, others to do work • LF lets threads take turns rotating between those roles
HSHA and LF Context • A concurrent system with asynchronous events • Performs both synchronous and asynchronous services • Synchronous: copy data from one container to another • Asynchronous: socket or input stream reads and writes • Both kinds of services must interact • Thread share event sources • A thread may handle events for other threads • A thread may also handle its own events • Efficient processing of events is important
HSHA and LF Design Problem • Asynchronous processing to make a system responsive • E.g., dedicated input thread (or reactive socket handling) • Services may map directly to asynchronous mechanisms • E.g., hardware interrupts, signals, asynchronous I/O, etc. • Synchronous processing may be simpler, easier to design, implement, and (especially) debug • How to bring these paradigms together? • How to achieve high performance multi-threading? • Service requests arrive from many sources • Concurrency overhead must be minimized • Context switches, locking and unlocking, copying, etc. • Threads must avoid race conditions for event sources • And for messages, buffers, and other event processing artifacts
HSHA Solution • Decompose architecture into two service layers • Synchronous • Asynchronous • Add a queueing layer between them to facilitate communication event notification computation filtering classification (short duration operations) (long duration operations) Queue get (put) put (get) synchronous layer asynchronous layer
LF Solution • Threads in a pool take turns accessing event sources • Waiting threads (followers) queue up for work • Thread whose turn it is acts as the “leader” • May dispatch events to other appropriate (passive) objects • May hand off events to other threads (e.g., active objects), i.e., “sorting the mail” until it finds its own work to do • Leader thread eventually takes event and processes it • At which point it leaves the event source and … • … when it’s done processing becomes a follower again … • … but meanwhile, another thread becomes leader • Protocol for choosing and activating a new leader • Can be sophisticated (queue w/ cond_t) or simple (mutex) • Should be appropriately efficient
Example: “Half-Sync/Half-Reactive” • HSHA Variant • Notice asynchronous, queuing, synchronous layers (HSHA) • Generalizes dedicated input thread to multi-connection case • Easy to multi-thread synchronous layer as shown below
Limitations of the HS/HR Approach • Data passing overhead • I/O thread to Worker thread • Dynamic memory allocation • Synchronization operations • Blocking factors • Context switches • May see unnecessary latency • Best case and common case may differ if work is heterogeneous • May suffer more cache misses • Due to multiple threads touching the same data
Example Revised: Applying LF Pattern • Allocate a pool of application threads, as with HSHR • However, don’t separate synchronous/reactive layers • Allow threads from the pool down into the reactor • Threads from pool take turns being the leader • Leader gets an event • Looks at its destination info • ACT, handle, etc. • If for leader, just handle it • Leader leaves reactive layer • Threads “elect” a new leader • Leader enters reactor • Other threads block on CV(s) • Otherwise, leader queues it • Notify queue CV(s) • Thread(s) wake up, access queue, get their events, handle them Leader